The latest articles on DEV Community by Hannan Zubair (@hannanthetron). https://dev.to/hannanthetron https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4192939%2F17385929-f2ea-43c6-b2a5-b37468053562.png https://dev.to/hannanthetron en Hannan Zubair Mon, 26 Oct 2026 05:18:22 +0000 https://dev.to/hannanthetron/scamsloth-listens-to-your-voicemails-so-you-dont-have-to-5cbi https://dev.to/hannanthetron/scamsloth-listens-to-your-voicemails-so-you-dont-have-to-5cbi <p><em>This is a submission for the <a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19">Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady</a></em></p> <h2> What I Built </h2> <p>I never listen to my voicemails. Half of them are scams, and the other half I worry might be. ScamSloth does the listening for me. Upload a voicemail and in about 4 seconds you get a <strong>Scam / Not scam</strong> verdict with red flags, a check on whether the <strong>callback number</strong> is reported as a scam online, and the full transcript. No listening and no googling numbers.</p> <h2> Demo </h2> <p>Live: <a href="https://scamslothis.mildlyintelligent.tech" rel="noopener noreferrer">https://scamslothis.mildlyintelligent.tech</a></p> <h2> Code </h2> <p><a href="https://github.com/bauji/scamslothis" rel="noopener noreferrer">https://github.com/bauji/scamslothis</a></p> <h2> How I Built It </h2> <ul> <li> <strong>Qwen3.5-4B (open-weight)</strong> on <strong>Tinker</strong> is the brain. It reads the transcript and returns a JSON verdict with up to 5 red flags. Qwen3.5-4B is released under Apache 2.0, so I'm free to fine-tune it, self-host it, and ship my own version.</li> <li> <strong>ElevenLabs Scribe</strong> turns the audio into text.</li> <li> <strong>SerpApi</strong> searches the callback number for scam reports. ScamSloth pulls the number out of the transcript even when it's spoken as words ("four one six…").</li> <li> <strong>FastAPI</strong>, managed with <strong>uv</strong>, deployed on <strong>Render</strong> at my own <code>.tech</code> subdomain.</li> </ul> <p>I built it in small steps, each one leaving a working app: first a page, then the transcript, then the verdict, then the number check.</p> <h2> Why Does Open Innovation Matter? </h2> <p>A scam checker has to know about <em>your</em> life. While building ScamSloth, I tested it on my own voicemails, and it flagged one of my legitimate ones as a scam. The base model (Qwen3.5-4B) had no context on my life. If it had known my work history and which financial institutions I use, it wouldn't have flagged that call. The funny part is that the exact same voicemail sent to <em>you</em> could very well be a scam.</p> <p>Generic models are great at a lot of things, but they don't know my data or my life. I see a future where intelligence works for everyone: open models with open weights, fine-tuned on each person's own data. Everyone gets intelligence customized for them, and without ads 😉.</p> <p>To be upfront: transcription (ElevenLabs) and the number search (SerpApi) are closed services today. The roadmap swaps in open speech-to-text and local inference.</p> <h2> My Agent Session </h2> <h2> The Lazy Payoff </h2> <p>I ran it on 4 of my own voicemails this week. Each took about 4 seconds instead of listening and googling the number. It caught both spam calls, but it also flagged one real call, and that's exactly the mistake my fine-tune is going to fix.</p> <h2> Prize Categories </h2> <ul> <li> <strong>Best Use of Render</strong>: the whole app (FastAPI front end and AI pipeline) runs on Render.</li> <li> <strong>Best Use of ElevenLabs</strong>: Scribe transcribes the voicemail audio for the open-weight model.</li> <li> <strong>Best Use of SerpApi</strong>: live web search checks whether the callback number is reported as a scam.</li> <li> <strong>Best Use of a .tech Domain</strong>: live at scamslothis.mildlyintelligent.tech.</li> </ul> <h2> What's Next </h2> <ul> <li>Fine-tune on my own labelled voicemails and score it against the base model</li> <li>Turn it into an agent that decides when to look up a number</li> <li>Open speech-to-text and local inference</li> </ul> devchallenge hf26challenge